Ahead of the holidays, Microsoft said it was upgrading the AI model behind Bing Image Creator, the AI-powered image editing tool built into the company’s Bing search engine. Microsoft promised that the new model — the latest version of OpenAI’s DALL-E 3 model, code-named “PR16” — would allow users to create images “twice as fast as before” with “higher quality.”

But it didn’t deliver. Complaints quickly flooded X and Reddit.

“The DALL-E we used to love is gone forever,” said one Redditor. “I’m using ChatGPT now because Bing has become useless for me,” wrote another.

The blowback was such that Microsoft said it’ll restore the previous model to Bing Image Creator until it can address the issues.

bring back the old dalle 3! the image quality is so much better on the old model. like these images for example. the image generated by the new model sucks 🙁 pic.twitter.com/BjIM8MS4ng

— ze ᡣ𐭩ྀིྀི (@riegrowl) December 28, 2024

“We’ve been able to [reproduce] some of the issues reported, and plan to revert to [DALL-E 3] PR13 until we can fix them,” Jordi Ribas, head of search at Microsoft, said in a post on X Tuesday evening. “The deployment process is very slow unfortunately. It started over a week ago and will take 2-3 more weeks to get to 100%.”

So what went wrong?

It’s difficult to compare model outputs from anecdotal reports, particularly when the prompts aren’t standardized. But many users said that PR16 tended to make images look less realistic — and “lifeless.” Mayank Parmar, writing for Windows Latest, noted that PR16-generated images lacked detail and polish, and appeared weirdly cartoonish.

I don’t know who you think you’re kidding with this. DALL-E is objectively worse than it ever was after this “update” and you’re being outpaced by other companies like Google. It’s absolutely night and day comparing image quality now to just a couple months ago pic.twitter.com/EdSdk7aign

— outward (@roccynoxy) December 19, 2024

It’s not the first time an image model that presumably passed internal checks wasn’t well received publicly. Back in February, Google was forced to pause its AI chatbot Gemini’s ability to create images of people after users complained of historical inaccuracies.

The missteps illustrate just how challenging it can be to measure model improvements in the real world. According to Ribas, Microsoft’s benchmarking found PR16’s quality to be “a bit better on average” compared to the previous Bing Image Creator model.

Whatever internal metric the company used, it seems clear that it didn’t align with most people’s preferences.